"I built a (very small) long running agent" — Analytics Engineering Roundup (dbt Labs)

Why this is in the vault

A dbt Labs essay builds a minimal long-running agent ("Tinyberg") with Claude Code's goal loop and argues that verifiability is the gating constraint on which data-team jobs are tractable today — directly on-point for RDCO's background-agents architecture work happening this same day.

The core argument

The author follows up a prior dbt Roundup piece (Tristan Handy's "agent swarm" experiment) by testing a different emerging modality: the long-running agent. The framing sketches a progression of agent patterns now in use by data teams — one-off triggered tasks (the "analytics agent Slack channel"), paired human+agent development (Claude Code, the dbt Developer Agent) — and then the frontier patterns only now appearing: multiagent monitoring systems, and long-running agents capable of much larger goals like "migrate my entire dbt project to Iceberg" or "refactor from Star Schema to Data Vault."

The author cites Cursor's from-scratch browser build and Anthropic's clean-room C compiler as proof that agents can handle far more complexity than commonly assumed, with one heavy caveat: both are extremely verifiable tasks where the agent can check itself against an external oracle. The "Ralph Loop" (give an agent a goal, let it iterate until it verifies completion) is named as the popular minimal variant, now surfacing as a built-in goal command in major harnesses.

To demonstrate, the author built Tinyberg: a deliberately small, read-only Apache Iceberg table inspector. Workflow: co-write a spec with GPT-5.5, drop spec.md into a directory, boot Claude Code with goal, and "let it rip." After ~2 hours unattended, the agent shipped a working implementation hitting all six spec requirements (load a local Iceberg table, resolve snapshots, read manifest metadata, produce scan plans, a CLI, and a test suite) — with the constraint that mature engines (DuckDB, PyIceberg) could appear only in tests as an oracle, never in Tinyberg's runtime.

The load-bearing claim: verifiability determines deployability. Good near-term candidates all sit at the intersection of highly-verifiable and valuable — platform migrations (validate data matches exactly), dbt best-practices enforcement (iterate until conformant to dbt project evaluator), library/engine upgrades (e.g. to the dbt Fusion engine), pipeline spend/speed optimization, and data-diff remediation ("iterate until the diff is gone"). Poor candidates are precisely the ones lacking ground truth: building a semantic layer from scratch (its value is organizational metric buy-in, not in the verification loop) or deciding data-platform architecture (architecture decisions become the ground truth, so there's nothing external to verify against).

Looking forward, the author points to Anthropic's "effective harnesses for long-running agents" guidance: the next step beyond a single goaled loop is multiple agents sharing a filesystem but operating in separate context windows, with a dedicated evaluator agent that dramatically improves performance even on outcomes that aren't fully verifiable from scratch. The piece closes expecting forward-looking data teams to operationalize this soon (with a soft CTA to bring cool implementations to the dbt Summit).

Mapping against Ray Data Co

• Strong, same-day relevance to the background-agents build. RDCO is literally offloading scheduled loops into background agents today; this essay is the data-engineering-flavored statement of the same architecture. The progression it describes (triggered one-offs → paired dev → long-running goaled loops → multi-agent shared-filesystem-with-evaluator) is exactly the trajectory the harness-thesis cluster tracks. Pairs tightly with [[2026-05-08-dan-farrelly-background-agents-orchestration]].
• "Verifiability = deployability" validates RDCO's verification-as-independent-worker discipline. The author's frontier — an evaluator agent lifting performance on partially-verifiable outcomes — is the same move RDCO already codified for non-visual surfaces (/verify-vault-write, /verify-strategic-output, /verify-dispatch; SOP 2026-05-19-verification-as-independent-worker-pattern). RDCO should read its own fresh-eyes critic agents (design-critic, video-critic, verify-pdf-output) as exactly the "evaluator agent in a separate context window" this essay points to as the next step. We're ahead of the curve here, not behind.
• The verifiable-task list is a literal RDCO services/skills backlog. Data platform migrations, dbt best-practices enforcement, dbt-version/Fusion upgrades, pipeline cost/speed optimization, and data-diff remediation are precisely the highly-verifiable, valuable jobs RDCO's agent-deployer / "fractional forward-deployed engineer" wedge targets. The audit-model + generate-tests skills already live in this lane; a long-running "iterate until conformant" loop over a dbt project is a plausible next skill. Maps to the augmentation narrative in [[2026-04-04-claude-code-not-replacing-data-engineers]].
• Confirms dbt Labs as a committed agent-deployer (competitive/landscape read). The Roundup is dbt Labs' channel; this piece, plus the Developer Agent and ADE references, reinforce that dbt is shipping agent infrastructure into the same data-team surface RDCO eyes. See [[2026-04-15-dbt-ade-bench-data-agent-benchmark-stancil]].
• The "Ralph Loop" / goal-command framing is already in the vault. Worth treating this as corroboration rather than new signal; the agent-threads vocabulary (Ralph Wiggum loop, thread thickness) was captured in [[2026-04-20-indydevdan-agent-threads-boris-cherny]]. The new data point is its arrival as a first-class harness command and the verifiability filter for which loops to run.
• Caution on over-reading the demo. Tinyberg is an explicitly minimal, ~2-hour proof of concept on a maximally-verifiable read-only spec. Don't extrapolate "2 hours unattended ships production data infra." The honest constraint (oracle in tests only, read-only scope) is the load-bearing detail.

Related

• [[2026-05-08-dan-farrelly-background-agents-orchestration]]
• [[2026-05-20-arxiv-code-as-agent-harness]]
• [[2026-04-20-indydevdan-agent-threads-boris-cherny]]
• [[2026-04-15-dbt-ade-bench-data-agent-benchmark-stancil]]
• [[2026-04-04-claude-code-not-replacing-data-engineers]]